EDAI - AI-Assisted Data Explorer

⚠️ Important Note: This repository contains only the backend of the EDAI project. For the complete system functionality, you also need to set up the frontend, available at: EDAI Frontend Repository

---

🚀 Website Deployment

You can see the project up and running at the following link:
🔗 View Demo

---

📸 Screenshots

Generated data visualization

Visualization zoom

API view

---

📋 Description

EDAI (AI-Assisted Data Explorer) is an innovative web platform that enables data analysis through natural language, eliminating the need for technical knowledge in programming or statistics. Developed as a Professional Integrative Project (PIP) at Universidad Nacional Arturo Jauretche (UNAJ).

The system uses Large Language Models (LLMs) in a multi-agent architecture to interpret queries, execute analysis, and automatically generate visualizations, democratizing access to exploratory data analysis.

---

🎯 Key Features

• ✅ Conversational analysis: Perform queries in natural language without writing code
• ✅ Automatic graph generation: Histograms, scatter plots, heat maps, clustermaps and more
• ✅ Multi-agent architecture: Two specialized LLM agents working coordinately
• ✅ Conversational memory: Remembers previous interactions and learns user patterns
• ✅ Adaptive strategy: Automatic selection between SQL queries and DataFrame processing
• ✅ Fallback system: Automatic error recovery with up to 3 retries
• ✅ Intelligent deduplication: Duplicate file detection using SHA-256 hashing
• ✅ Modern and responsive interface: Adaptable design for desktop, tablet and mobile with light/dark mode

---

🏗️ System Architecture

General Structure

EDAI/
├── src/                          # Backend (FastAPI + Python)
│   ├── api/                      # HTTP exposure layer
│   │   ├── routes/               # Endpoint controllers
│   │   │   ├── chat.py          # Conversational interaction endpoint
│   │   │   └── documents.py     # Document management endpoints
│   │   ├── schemas/              # Validation models (Pydantic)
│   │   └── main.py              # FastAPI initialization
│   ├── services/                 # Business logic
│   │   ├── chat_service.py      # Query processing
│   │   └── document_service.py  # File management
│   ├── nodes.py                  # LangGraph graph nodes
│   ├── graph.py                  # Execution flow definition
│   ├── state.py                  # Shared agent state
│   ├── tools.py                  # Analysis and visualization tools
│   ├── prompts.py                # Dynamic prompt construction
│   ├── memory.py                 # Conversational memory system
│   ├── database.py               # PostgreSQL management
│   ├── dataset_manager.py        # Dataset loading and storage
│   ├── multi_dataset.py          # Intelligent dataset selection
│   ├── checkpoints.py            # State persistence (PostgresSaver)
│   ├── config.py                 # Global configuration
│   └── utils.py                  # Auxiliary utilities
│
├── outputs/                      # Generated graphs (.png)
├── requirements.txt              # Python dependencies
└── README.md                     # This file

Multi-Agent Architecture

The system implements a hybrid architecture with two specialized LLM agents:

1. GPT-OSS-120B (via Groq)

• Role: Semantic understanding agent
• Function: Generates semantic descriptions of datasets and selects the most relevant dataset through semantic matching

2. Gemini 2.0 Flash (Google)

• Role: Executor and reasoning agent
• Function: Processes queries, selects tools, generates Python/SQL code and produces visualizations

---

🛠️ Technologies Used

• Python 3.9+: Main language
• FastAPI: High-performance asynchronous web framework
• LangChain: Integration with language models
• LangGraph: Multi-agent conversational flow orchestration
• PostgreSQL: Database for storage and checkpoints
• Pandas: Data manipulation and analysis
• Matplotlib/Seaborn: Visualization generation
• Psycopg: PostgreSQL connector
• Pydantic: Data validation
• Google Generative AI: Client for Gemini 2.5 Flash
• Groq: Client for GPT-OSS-120B

---

🚀 Installation and Configuration

Prerequisites

• Python 3.9 or higher
• Node.js 18+ and npm/pnpm
• PostgreSQL 15+
• API Keys:
  • Google AI Studio (Gemini)
  • Groq Console (GPT-OSS-120B)

1. Clone the Repository

git clone https://github.com/FernandoBlancoTFL/edai-llm-backend.git
cd edai-llm-backend

2. Configuration

2.1 Create Python virtual environment

python -m venv venv

# Windows
venv\Scripts\activate

# Linux/Mac
source venv/bin/activate

2.2 Install dependencies

pip install -r requirements.txt

2.3 Configure PostgreSQL

Create database:

CREATE DATABASE edai_db;
CREATE USER edai_user WITH PASSWORD 'your_secure_password';
GRANT ALL PRIVILEGES ON DATABASE edai_db TO edai_user;

2.4 Configure environment variables

The project includes a .env.example file as a template. Copy it and rename it to .env, then fill in your configuration values:

# Copy the example file
cp .env.example .env

Edit the .env file with your actual values:

# PostgreSQL Configuration
POSTGRES_HOST=localhost
POSTGRES_PORT=5432
POSTGRES_USER=edai_user
POSTGRES_PASSWORD=your_secure_password
POSTGRES_DB=edai_db

# External APIs
GOOGLE_API_KEY=your_gemini_key_here
GROQ_API_KEY=your_groq_key_here

# Cloudinary keys
CLOUDINARY_CLOUD_NAME=your_cloudinary_cloud_name_here
CLOUDINARY_API_KEY=your_cloudinary_key_here
CLOUDINARY_API_SECRET=your_cloudinary_api_secret_here

Note: Never commit the .env file to version control. Only .env.example should be tracked by git.

Running the Application

• For local development: configure environment variables in the .env file
• For production: environment variables are configured in the hosting platform

---

2.5 Activate virtual environment

# From project root
.\env\Scripts\activate

2.6 Run backend

# From project root
uvicorn src.api.main:app --reload --host 0.0.0.0 --port 8000

Backend will be available at: http://localhost:8000
Interactive documentation (Swagger): http://localhost:8000/docs

---

📖 User Guide

1. Data Loading

1. Click "Upload File" in the sidebar
2. Select CSV or Excel file (.xlsx, .xls)
3. Wait for confirmation: "File uploaded successfully"
4. The system automatically generates a semantic description of the dataset

Note: The system detects duplicate files using SHA-256 hashing and notifies the user.

2. Making Queries

a) General Interaction

- "hello"
- "what can you do?"
- "help"
- "thanks"

b) Dataset Queries

- "What are the columns of the crocodile dataset?"
- "Give me the first 5 records of the trips dataset"
- "Average Booking Value for UPI payments"
- "How many crocodiles weigh more than 100 kg?"
- "Describe the statistics of the solar radiation dataset"

c) Graph Generation

- "Generate a histogram of Booking Value"
- "Scatter plot between num_hora and Rad_Solar"
- "Show correlation clustermap of the dataset"
- "Scatter plot of Rad_Solar vs num_dia with colors by Clarity Index"
- "Pie chart of Payment Method distribution"

3. Advanced Features

• Conversational memory: The system remembers previous interactions and learns usage patterns
• Automatic dataset selection: If you have multiple files loaded, the system identifies which is relevant for your query
• Intelligent fallback: If an SQL strategy fails, it automatically retries with DataFrames
• Graph zoom: Click on any generated graph to enlarge it

---

👥 Author

Fernando Jose Blanco
📧 fernando.blanco004@gmail.com
🎓 Computer Engineering - Universidad Nacional Arturo Jauretche (UNAJ)

Academic Supervision

• Organizational Tutor: Prof. Dr. Marcelo Ángel Cappelletti
• Supervisor Professor: Prof. Eng. Lucas Maximiliano Olivera
• Program Coordinator: Dr. Eng. Morales, Martín

---

📄 License

This project was developed as a Professional Integrative Project (PIP) at Universidad Nacional Arturo Jauretche (UNAJ) under the Student Research Initiation Scholarship (BIEI) program.

---

⭐ If you found this project useful, consider giving it a star on GitHub ⭐